Conference of neural Binding in space and time
March (15-18) - 2000 - Leibzig (Germany).
 
Object recognition using spiking neurons I: A model for rapid processing of natural images
 
Arnaud Delorme, Rufin van Rullen & Simon Thorpe
 

Centre de Recherche Cerveau et Cognition, CNRS, Université Paul Sabatier, 133 route de Narbonne, F-31062 Toulouse, France

arno@cerco.ups-tlse.fr, rufin@cerco.ups-tlse.fr, thorpe@cerco.ups-tlse.fr

 
    Recent experiments concerning the speed at which targets in briefly flashed photographs of natural scenes can be detected [Thorpe et al, 1996 Nature (London) 381 520; Fabre-Thorpe et al, 1998 Neuroreport 9 303] impose strong constraints on models of processing in the visual system. We propose a biologically plausible model for object recognition consistent with this sort of rapid processing that involves a hierarchically organised system of asynchronously discharging integrate-and-fire neurons. Starting with the retina, the earliest firing cells are those with the strongest inputs. At subsequent stages, more complex receptive field properties, such as those observed in V1, result from integration of these input spikes and lateral interactions. The final processing layers contain very large numbers of neurons trained to particular views of the various targets. Simulations using SpikeNET, a software system designed for modelling networks with millions of integrate-and-fire neurons and billions of synapses, demonstrate that architectures based on these principles can indeed be used to perform sophisticated object recognition with natural scenes. If sufficient neurons are available, it is possible to design networks of this type that are capable of simultaneously recognising and localising an arbitrary number of objects in the visual field, thus avoiding the problem of binding identity with location.